Publication number | US7523117 B2 |
Publication type | Grant |
Application number | US 11/416,766 |
Publication date | Apr 21, 2009 |
Filing date | May 3, 2006 |
Priority date | May 4, 2005 |
Fee status | Paid |
Also published as | US20060274062, WO2006119482A2, WO2006119482A3 |
Publication number | 11416766, 416766, US 7523117 B2, US 7523117B2, US-B2-7523117, US7523117 B2, US7523117B2 |
Inventors | Cun-Quan Zhang, Yongbin Ou |
Original Assignee | West Virginia University Research Corporation |
Export Citation | BiBTeX, EndNote, RefMan |
Patent Citations (36), Non-Patent Citations (3), Referenced by (7), Classifications (13), Legal Events (5) | |
External Links: USPTO, USPTO Assignment, Espacenet | |
This application claims priority to U.S. Provisional Patent Application Ser. No. 60/677,655, filed May 4, 2005, the entire content of which is incorporated herein by reference.
Research carried out in connection with this invention was supported in part by National Security Agency Grant Nos. MDA904-01-1-0022 and MSPR-03G-023. Accordingly, the United States government may have certain rights in the invention.
This application includes a computer program listing appendix on two duplicate compact discs. Each compact disc includes a single file named “clustering.txt,” created May 1, 2006. The program listing is in C++ language and the size of the file is 57 kilobytes. The contents of the computer program listing appendix are hereby incorporated herein by reference.
The present invention relates to methods, processes and systems for working with and analyzing data sets and, more specifically, for clustering a data set into subsets of closely related objects.
Many applications require partitions of a large graph/network into smaller communities. Qualitatively, a community is defined as a subset of vertices within the graph such that connections between the vertices are denser than connections with the rest of the network. The detection of the community structure in a network is generally intended as a procedure for mapping the network into a tree. In this tree (called a dendrogram in the social sciences, or a hierarchical tree in biology), the leaves are the vertices whereas the edges join vertices or communities (groups of vertices), thus identifying a hierarchical structure of communities nested within each other.
Partitioning graphs into communities and searching for subgraphs with high internal density within graphs/networks is of practical use in various fields: parallel computing, the Internet, biology, social systems, traffic management, etc.
For example, the following is an application in biology: Complex cellular processes are modular, that is they are accomplished by the concerted action of functional modules. These modules are made up of groups of genes or proteins involved in common elementary biological functions. One important and largely unsolved goal of functional genomics is the identification of functional modules from genomewide information, such as transcription profiles or protein interactions. To cope with the ever-increasing volume and complexity of protein interaction data, new automated approaches for pattern discovery in these densely connected interaction networks are required. Cluster analysis is an obvious choice of methodology for the extraction of functional modules from protein interaction networks. (See Detection of Functional Modules From Protein Interaction Networks, by Pereira-Leal, etc., Proteins, 2004; 54:49-57.)
A second example comes from the study of social networks: It is widely assumed that most social networks show “community structure”, i.e., groups of vertices that have a high density of edges within them, with a lower density of edges between groups. It is a matter of common experience that people divide into groups along lines of interest, occupation, age, etc. (See The structure and function of complex networks, By Newman, SIAM Review 45, 2003; 167-256)
Due to the fact of its importance in applications, many clustering methods/algorithms have been discovered and patented (such as U.S. Pat. No. 5,040,133 Feintuch, et al., U.S. Pat. No. 5,263,120 Bickel, U.S. Pat. No. 5,555,196 Asano, U.S. Pat. No. 5,703,959 Asano, et al., U.S. Pat. No. 5,745,749 Onodera, U.S. Pat. No. 5,832,182 Zhang, et al., U.S. Pat. No. 5,864,845 Voorhees, et al., U.S. Pat. No. 5,940,832 Hamada, et al., U.S. Pat. No. 6,003,029 Agrawal, et al., U.S. Pat. No. 6,038,557 Silverstein, U.S. Pat. No. 6,049,797 Guha, et al., U.S. Pat. No. 6,092,072 Guha, et al, U.S. Pat. No. 6,134,541 Castelli, et al., U.S. Pat. No. 6,195,659 Hyatt, U.S. Pat. No. 6,269,376 Dhillon, et al., U.S. Pat. No. 6,353,832 Acharya, et al., U.S. Pat. No. 6,381,605 Kothuri, et al, U.S. Pat. No. 6,397,166 Leung, et al., U.S. Pat. No. 6,466,946 Mishra, et al., U.S. Pat. No. 6,487,546 Witkowski, U.S. Pat. No. 6,505,205 Kothuri, et al, U.S. Pat. No. 6,584,456 Dom, et al., U.S. Pat. No. 6,640,227 Andreev, U.S. Pat. No. 6,643,629 Ramaswamy, et al., U.S. Pat. No. 6,684,177 Mishra, et al., U.S. Pat. No. 6,728,715 Astley, et al., U.S. Pat. No. 6,751,621 Calistri-Yeh, et al., U.S. Pat. No. 6,829,561 Keller, et al., etc.), and some algorithms have been embedded in various popular software (such as, BMDP, SAS, SPSS-X, CLUSTAN, MICRO-CLUSTER, ALLOC, IMSL, NT-, NTSYS-pc, etc.).
In general, almost all existing clustering methods can be classified as one of two types: agglomerative or divisive, depending on how the hierarchical trees are constructed and how vertices are grouped together into communities. (Examples of agglomerative clustering algorithm are found in U.S. Pat. No. 5,040,133 Feintuch, et al., U.S. Pat. No. 5,832,182 Zhang, et al., U.S. Pat. No. 6,049,797 Guha, et al., U.S. Pat. No. 6,092,072 Guha, et al, U.S. Pat. No. 6,134,541 Castelli, et al., U.S. Pat. No. 6,195,659 Hyatt, U.S. Pat. No. 6,397,166 Leung, et al., etc. Examples of divisive clustering algorithm are found in U.S. Pat. No. 6,038,557 Silverstein, U.S. Pat. No. 6,353,832 Acharya, et al., U.S. Pat. No. 6,381,605 Kothuri, et al, U.S. Pat. No. 6,466,946 Mishra, et al., U.S. Pat. No. 6,505,205 Kothuri, et al, U.S. Pat. No. 6,640,227 Andreev, U.S. Pat. No. 6,684,177 Mishra, et al.; etc.)
The present invention provides a computer based method and a system for working with and analyzing data sets and, more specifically, for clustering a data set into subsets of closely related objects. The “objects” may be physical objects, such as people, genes or proteins, or portions of physical objects, or may represent less tangible data. A “level of relatedness” between objects of interest may represent any type of relationship between the “objects”, such as how closely related they are or how similar they are. Alternatively, the level of relatedness may represent how dissimilar objects are, depending on the application.
The present invention is a computer based method, or may take the form of a general or special purpose computing system that implements the method. The invention may also take the form of a computer readable medium having computer-executable instructions embodied therein for performing the computer based method. Data to be processed by the method or system may be entered by a user, provided as an output from another device or system, or may be stored on a computer readable medium or computing device.
Some embodiments of the computer based method make use of a practical (polynomial) algorithm to detect all subgraphs whose dynamic density is k (for a given integer k). The algorithm, uses a well-defined measure of density and achieves its goal optimally; that is, it finds exactly the optimal solution, not just an approximation.
According to a first embodiment of the present invention, a computer-based method of clustering related data is provided. The data represents a plurality of objects of interest and information about levels of relatedness between pairs of the objects. The computer based method comprises:
In some versions, the level of relatedness between objects of interest represents a similarity or closeness between the objects of interest.
In the computer based method, finding all possible subgraphs H of G may be accomplished by finding the maximal subgraph H that, for every edge e, H−e contains k edge-disjoint spanning trees.
Preferably, G is treated as the only level-0 community and finding all possible subgraphs H of G is accomplished by finding all level-k communities within a previously found level-(k−1) community, and repeating this finding step for k←k+1 until only single vertices remain. Finding all level-k communities within a level-(k−1) community H may be accomplished by:
a) letting T_{1}, T_{2}, . . . , T_{k−1 }be edge-disjoint spanning trees of H;
b) finding a spanning forest T_{k }in
c) finding an edge e that is not used up in the set of T_{i }for all i=1, . . . , k, the edge being a seed edge;
d) establishing an edge subset B_{p}, starting with p=1, which initially contains the seed edge e;
e) expanding the subset B_{p}, recursively, for each T_{i }and each e′∈ B_{p}, by adding all edges e* of any circuit in T_{i}+e′;
f) repeating step (e) until either;
g) if Case 1 of step (f) occurs, adjusting the set of spanning forests {T_{1}, T_{2}, . . . , T_{k}} and expanding the spanning forest T_{k }and, thereafter, repeating step (c) for the adjusted set of forests {T_{1}, T_{2}, . . . , T_{k}};
h) if Case 2 of step (f) occurs, storing the subset B_{p }and setting p←p+1 and repeating step (c) with an edge e that also does not join the same vertices as any edge in any of B_{1}, B_{2}, . . . B_{p−1};
i) merging B_{p}S that overlap; and
j) outputting the set of subgraphs induced by stored subsets B_{p }resulting from step (i), each of which is a level-k community of G and contained in H.
Expanding the forest T_{k }as required in step (g) may be accomplished by recording a track of replacement for every edge of B_{p}, and then adjusting the set of forests {T_{1}, T_{2}, . . . , T_{k}} by adding the seed edge, thereby expanding the spanning forest T_{k }by connecting unconnected portions of T_{k}. Recording the track of replacement for every edge of B_{p }may be accomplished by recording the track of replacement for the seed edge e by initializing sequences I(e)=Ø and S(e)={e}, and recording the track of replacement for each edge e* in the circuit of T_{i}+e′ by sequences I(e*)=I(e′)i and S(e*)=S(e′)e′.
Expanding the forest T_{k }as required in step (g) may include letting e′ be the edge of B_{p }joining two unconnected portions of T_{k }and letting I(e′)=i_{1}i_{2 }. . . i_{h−1 }and S(e′)=e_{1}e_{2 }. . . e_{h }where e_{h}=the seed edge, and setting T_{k}←T_{k}+e′ and for each r=1, . . . , h−1, setting T_{i} _{ r }←T_{i} _{ r }+e_{r}−e_{r+1}.
According to another embodiment of the present invention, a system is provided for determining a level of relatedness of data within a dataset. The system includes a computer processor, a memory in communication with the processor, an output device in communication with the processor, and a computer readable medium having computer-executable instructions embodied therein. The computer executable instructions perform a method comprising:
The processor is operable to execute the computer-executable instructions embodied on the computer readable medium.
In yet another embodiment of the present invention, a computer readable medium is provided, having computer-executable instructions embodied therein for performing a method of clustering related data representing a plurality of objects of interest and information about levels of relatedness between pairs of the objects. The method comprises:
The above described embodiments can be further revised to achieve an improved complexity, as will be clear to those of skill in the art, based on a review of the following. Other embodiments and variations on the present invention will also be clear to those of skill in the art based on a review of the following specification and Figures.
General Overview
Some embodiments of the present invention provide improved methods, systems and algorithms for data clustering. These embodiments use mathematical graph theory techniques to provide a practical method of clustering a large dataset into subsets of the most closely connected objects. That is, when a dataset is made up of a network of input objects (such as the inhabitants of a town) that have associated levels of connection between them (such as the number of interactions between two inhabitants), the algorithm clusters these objects into subsets such that those subsets are interconnected to a proscribed degree and are as inclusive as possible. Embodiments of the present invention accomplish this by considering the objects as vertices in a weighted graph (in the graph theory sense) and the level of connectivity between a pair of objects as the number of edges connecting those vertices (also called the weight of the edge). Specifically, it considers the problem in terms of building trees spanning the graph (connecting every point with no unnecessary edges) and then, by means of unsaturated/used edges (that is, edges whose multiplicity/weight is greater than the number of times it appears in the set of trees), building the clusters by collecting well-connected edges (and, by extension, building a cluster of their associated vertices/objects) in a way that is computationally practical and also produces the unique cluster sets of a given connectedness. This is done such that the result is independent of any ordering imposed on the data in the input process.
More specifically, embodiments of the present invention rely on the concept of dynamic density to define connectedness within a data subset (that is, the associated subgraph). Though dynamic density is based on a definition that considers all possible ways to further partition the subset in question, embodiments of the present invention are able to build both the trees and the cluster sets iteratively with a guaranteed minimum dynamic density in an efficient manner. These embodiments do this by building and maintaining a set of edge-disjoint trees in such a way that the number of such trees spanning a subset while leaving at least one unsaturated edge can be used as a measure of the connectivity (that is, it is equivalent to the dynamic density). The unsaturated edges, that is those edges not being used by any spanning tree, are then utilized both to build the next tree within a currently considered set (if possible) and as starting points for finding internal data clusters of a higher level of connectivity than at the previous iteration of the process.
The process of the present invention is able to manipulate trees efficiently because it has a “fixed reference frame”. That is to say, even as the algorithm searches through edges for ones to include in the potential cluster under construction, adjusting the trees as it goes, the set of new edges to be considered for inclusion is not affected by these necessary adjustments of the trees that are central to the process.
Simple Example to Illustrate the General Concept
To explain the main features of the method in less complex mathematical language, one might say the method of this invention is a method for partitioning a dataset made up of a set of items and a set of quantified relations between those items, each relation being a positive integer (1, 2, 3, . . . ). A small example would be the items {a,b,c} and relations {ab} weight 2, {ac} weight 5, {bc} weight 1.
To see what partitioning by the number of connections looks like, a slightly more complicated graph is useful.
As stated previously, this invention makes use of a list of spanning trees to calculate the connectedness of a graph. A spanning tree is a subgraph (a graph that is a subset of the main graph's edges and vertices) that connects all vertices of the graph to be spanned using a minimum number of edges, that is one less than the number of vertices. An alternative definition of a spanning tree is a subgraph that connects all the vertices of a graph but contains no circuit (a circuit can be thought of as two paths between the same pair of vertices).
The spanning tree on the left is not unique. Instead of the edge {ab}, {ac} could have been used to connect {a} to the rest of the vertices. Thus a particular spanning tree or set of spanning trees might turn out to be suboptimal. For this reason, the present invention adjusts trees in the course of the algorithm. It does this, however, in such a way as to not increase the computational complexity to the point of making the calculation unfeasible; in particular, it does not search through all possible sets of spanning trees, but yet it produces an optimal solution independent of what method is used to make the trees (in fact, the algorithm can build a tree “from scratch”, but there are a number of standard algorithms that can be used to efficiently start the process).
The example of
For the example of
Explanation of Formal Mathematical Basis for Invention
We turn now to an explanation of the formal mathematical underpinning of the present invention. This will be followed with an algorithm that represents one embodiment of the present invention, described in formal mathematical terms. The same embodiment will then be explained in more detail, in something closer to plain English. An example using the starting dataset of
As a prelude to the formal mathematical description, it is necessary to define both the symbols used and the mathematical underpinning so that a precise description can be made.
Weight Graphs
In this invention, an input of a collection of related data is presented as a graph model G with a vertex set V(G) and an edge set E(G). Each vertex v of the graph G represents an input object. Each edge e of the graph G is a connection between some pair of vertices of V(G). This edge e represents the closeness/similarity of the two objects in the input data associated with the vertices. For the case of a weighted graph, which is the sort frequently encountered in practical applications, each edge e has associated with it an integer weight w(e). This weight can be thought of as representing w(e) parallel edges connecting the pair of vertices in the graph G.
Dynamic Density
Let H be a subgraph of graph G. The dynamic density of H is the greatest integer k such that
where the minimum is taken over all possible partitions P of the vertex set of H, and E(H/P) is the set of crossing edges between parts of P.
To explain again for those not fully familiar with this mathematical notation, H, as a subgraph of G, is itself also a graph, which is a collection of vertices V(H) and edges E(H). The vertex set, V(H), can be partitioned as P=(V_{1}, V_{2}, . . . , V_{t}) (where each V_{i }represents a subset of vertices) and the set of edges in E(H) connecting vertices of H in different V_{i }(as opposed to those connecting two vertices in the same V_{i}) is written E(H/P). Said another way, H/P is a graph where each vertex corresponds to some set of vertices V_{i }of V(H) and each edge corresponds to an edge of H that crosses between two such sets. The number of edges (counting an edge e with weight w(e) as w(e) edges) is |E(H/P)|. This is normalized to take into account the size of P by dividing the number of crossing edges by |P|−1, where |P| is the number of sets into which P divides V(H).
The dynamic density is a well-defined way to quantify how internally connected a subgraph H is. If k is very high, then all parts of H are closely connected: no matter how one subdivides H, those subdivisions are closely connected to each other. If dynamic density k is low, some parts of H are only loosely connected to each other.
A maximal connected subgraph H of G with dynamic density at least k is a level-k community. By maximal here, we mean as inclusive as possible both of vertices from V(G) (every vertex that can be included is included) and also of edges (the edges E(H) are exactly those edges in E(G) that connect vertices V(H)). So in terms of this definition, the goal of this invention is, for a given integer h and an input graph G, to find a partition {V_{1}, V_{2}, . . . , V_{t}} of V(G) such that each subgraph of G induced by V_{i }is a level-h community. That is, every subgraph of G induced by V_{i }is of dynamic density at least h, while every subgraph that contains some vertices from two distinct parts V_{i }and V_{j }is of dynamic density less than h.
The full output of the algorithm is a hierarchical tree with the entire input graph G as the root (the level-0 community) constructed so that each node N in the k-th level of the tree represents a level-k community of the input G with its children (in the (k+1)th-level of the tree) all the level-(k+1) communities contained in N. The goal as stated in the previous paragraph would represent the hth level of this hierarchical tree of graphs.
Background of the Mathematical Issues Associated with this Invention
The most critical issue for algorithm design is complexity.
The number of partitions of a set of order n is call the Bell number and is denoted by B_{n }(see the book by P. J. Cameron (1994) “Combinatorics: topics, techniques, algorithms” p39, and the book by G. E. Andrews (1984) “The Theory of Partitions” p214). It can be shown that the Bell number B, satisfies the relation
B_{n}>2^{n−1}
by applying recursively that the number of partitions of a set of size n into exactly two non-empty subsets is 2^{n−1}−1. The above estimation of the Bell number can also be proved by applying the following recursive formula
(see the book by P. J. Cameron (1994) “Combinatorics: topics, techniques, algorithms” p 40) and that the sum of binomial coefficients is 2^{n−1 }and each B_{k−1}≧1.
So one can conclude that the Bell number B_{n }is at least an exponential function. That is, the complexity of the determination of the dynamic density of a graph would be unfeasible if it was determined by searching all possible partitions of the vertex set.
However, the algorithm presented herein is able to determine the dynamic density with a polynomial complexity (at most O(h^{2}n^{2}) and O(m^{2}), where n=|V(G)|, the number of vertices/data items, m=|E(G)|, the total weight of G, and h is the maximum dynamic density of the goal partition). That is, the algorithm is feasible and practical.
As was described earlier, this invention is based on a description using spanning trees which is equivalent to the one above in terms of dynamic density. It is this that allows a practical algorithm.
Mathematical Claim: A graph H is of dynamic density at least k if and only if, for every edge e of H, the subgraph H−e contains k edge-disjoint spanning trees.
The proof of this claim is presented below. First, it is appropriate to define again here a few important technical graph theory terms that will be needed here and in later items describing the algorithm.
Let e be an arbitrary edge of H. In the graph H−e, we have the following inequality for every partition P of V(H):
By a theorem in graph theory (Tutte and Nash-Williams, J. London Math. Soc. 1961), the graph H−e contains k edge-disjoint spanning trees. This completes one direction of the proof.
Next, we assume that, for every edge e of H, the graph H−e contains k edge-disjoint spanning trees T_{1}, . . . T_{k}. If there is a partition P of V(H) such that
Choose e as a cross edge between two parts of P. By contracting each part of P to a single vertex, each T_{i}/P becomes a connected, spanning subgraph of (H−e)/P, each of which contains at least |P|−1 edges (the number of cross edges in T_{1}/P). The sum of all those cross edges would contradict the above inequality; this completes the proof in the other direction.
Summary List of Main Variables Used in the Algorithm
The input graph G may have parallel edges. For each edge e, the multiplicity of e is denoted by w(e) (the number of parallel edges between a pair of given vertices). For a set T of spanning forests (that is, a subgraph of G connecting its vertices which contains no circuits), the coverage of an edge e is the number of members of T containing the edge e, denoted by c_{T}(e). (In the algorithm, it is required that c_{T}(e)≦w(e); that is, one can only use an edge w(e) times). The set of edges with c_{T}(e)≦w(e) (unsaturated edges) is denoted by E_{0}. These edges can be thought of as edges leftover from the edges of G after the construction of the set T of spanning forests.
The input graph G itself is considered as a level-0 community if E(G)≠Ø (where Ø indicates an empty set). E(G)=Ø represents the trivial case of no connections between vertices.
This example makes use of a “flag” to act as a “stop and restart” sign for the completion of the current community search and the starting point of searching another community. This flag is not “on” until the search of the current community ends. That is, the process of searching a new “community” is to start when the “flag” is “turned on”. In the following sample algorithm, the “flag” is a function ƒ associated with each e∈E(G) (where “∈” means “element of”). Initially all functions/flags ƒ(e) are set to 0. After an edge has been processed, its ƒ-value is set to 1.
This example also makes use of a “Track record”. A track record is a set of sequences (edge-sequence, index-sequence, or other sequences) associated with each edge (or vertex) of G. The purpose of a “track record” is to record the track of any possible adjustment and expansion of a set of spanning trees/forests. In this sample algorithm, the track record is a pair of sequences associated with each edge e: a sequence S(e) of edges and a sequence I(e) of indices of forests.
The integer p counts the number of potential communities. The subset B_{p }collects edges of the current level-k-community.
Algorithm
A sample algorithm according to a first embodiment of the present invention is shown in flowchart form in
Inputs: (see S0 a in flowchart) a graph G with w(e) as the multiplicity for each edge e and an integer h.
Goal: find all level-h-communities in G.
Referring to S0 b in the flowchart, the initial conditions are set. This step sets k=0 (the program runs until k=h level is complete), H←G (a level-0-community), T←Ø (the set of spanning trees in H) and c_{T}(e)=0 for all e in E(G).
Step 1 (see S1 a in flowchart)
Further Description of Algorithm
The following is a well-known lemma in mathematics, which plays a key role in the search for possible expansions of the last spanning forest T_{k }and the search for edges in the same community.
Let T be a spanning forest of a graph G and let e be an edge of E(G)−E(T) (an edge not contained in the forest T). Then, either T+e is a larger spanning forest, or T+e contains a unique circuit. Furthermore, if T+e contains a circuit C, then T+e−e′ is again a spanning forest of G for every edge e′ of the circuit C (in other words, e′ is any edge in the circuit created by adding e to T).
Communities are detected level-by-level in G. The main part of the algorithm is repeated for each k=1,2, . . . , h. For each k, the algorithm detects all level-k-communities in a previously identified level-(k−1)-community H. The following is an explanation of the main idea of the algorithm.
means union of all these edges sets, that is the union of all E(T_{1}), E(T_{2}), E(T_{3}), . . . , E(T_{k})).
The sample algorithm described shows the most basic and fundamental ideas of this invention. Using the same or similar ideas, the algorithm can be further revised or improved for the purpose of the reduction of time complexity. Different methods or approaches include (but are not limited to):
One preferred part of this invention is the connectivity measurement by the parking of spanning trees with an extra edge. An extra edge makes the search processing traceable, and therefore, practically feasible. Let e_{0}be an edge not contained in any of the forests T_{i}, for every i=1, . . . k. Let R(e_{0})←{e_{0}} be the set of edges that are replaceable by e_{0}. Add all edges e_{y }into R(e_{0}) for which e_{y }is contained in the circuit of T_{j}+e_{x }for some e_{x}∈ R(e_{0}). Elements of R(e_{0}) are edges replaceable by e_{0}. The subgraph of G induced by R(e_{0}) is a subgraph with dynamic density k whenever the spanning forests have reached their maximum.
The existence of an extra edge in a community eliminates some less connected outputs.
Fixed Reference Frame—an Important Method and Technique in the Algorithm Design that Reduces the Time-complexity and Memory Storage Complexity.
All edges that are replaceable by an extra edge e_{0}are determined by referring to the initial set of forests {T_{1}, . . . , T_{k}}. The correctness of the expansion procedure for the last spanning forest T_{k }and the searching procedure of all edges replaceable by e_{0}can be mathematically proved. Therefore, it eliminates the complexity of recording and tracing all modified forests for every additional edge in R(e_{0}).
The Classification of the Method—Neither Agglomerative Nor Divisive.
Almost all existing methods for clustering can be classified as two types: agglomerative and divisive, depends on the constructions of the hierarchical trees and the ways that vertices are grouped together into a community. However, the method presented in this patent is neither agglomerative nor divisive. Instead, the hierarchical tree is built from the top (the level-0 community—the input G itself) by dividing a level-(k−1)-community into several level-k-communities. And a k-community is detected by clustering vertices together based on their cyclic connection in a spanning tree of the level-(k−1)-community.
Special Features of the Method (About the Outputs)
Uniqueness of Outputs and Independence of Labeling
Output of a program according to embodiments of the present invention that searches communities with dynamic density as the measurement of density is independent of the labeling. That is, for two different labelings of the vertex set of the same input, the outputs of the program are the same. Note that most existing methods, either agglomerative or divisive, are not independent of labeling: computer programs follow the order of the labeling in each repeating or recursive computation and, therefore, vertices with smaller labels tend to have a higher priority of being selected. The outputs will be different if the input graph has some vertices with similar local structures which are close to each other. However, the output using the method of the algorithm of this patent is independent of the labeling and therefore the output of the same input is unique.
Density of Communities
For each level-k community H of G, the subgraph H is denser inside H than its connection with any other part outside of H. That is, the dynamic density of H is at least k, while any subgraph of G induced by some vertices of H and some vertices outside of H must be of dynamic density less than k.
Well-Defined Mathematically
The special features described in the previous two paragraphs are due to the fact that the measurement of density is mathematically well-defined, and the optimization goal is fully achieved by the program. Therefore, it is not an approximation approach.
Sample Algorithm in Plain Language
The above algorithm will now be described in more detail and in something closer to plain language.
Step 0: Initial Setup
All potential new communities have been built. These are B_{1}, B_{2}, . . . , B_{p}. These are sets of edges with associated vertices V(B_{i}). The construction guarantees that all vertices associated with a particular set have the required minimum connectedness. It does not guarantee that there is not overlap between these potential communities. It is necessary to systematically go through each pair of groups, B_{i }and B_{j }for i, j=1, 2, . . . , p with i≠j, and check whether they share any vertices. If so (V(B_{i})∩V(B_{j})≠Ø TRUE) then these two pair of communities will be merged. This continues until there is no overlap between the vertices. For convenience, renumber these so there are p′ communities and check for any: B_{1}, B_{2}, . . . , B_{p′}. These are the k-level communities of the graph H.
It should be noted that the above communities do not necessarily include all vertices of H. The missing vertices were not included because the clustering process of the algorithm is based on collecting edges. Thus single vertex k-level clusters are not represented. This is not a problem because (1) these singletons will remain singletons at all remaining levels, and so no further processing is necessary, and (2) given the starting dataset and an output made up only of k-level communities, it can be immediately inferred that objects missing from the communities are singleton clusters. While an implementation of this algorithm might add in the singleton clusters at some point for inclusion in the output, the implementation being described here does not.
CHECK LEVEL—Step 6a: called only from Step 5; followed by Step 6b or Step 6c (see below).
If the current level is not the final level, then each k-level community needs to be broken down into k+1-level communities. The final level is h, a level chosen by the user of the algorithm at the beginning. If k=h is TRUE (“YES”), then the algorithm proceeds to output the current communities (Step 6b). Otherwise (“NO”), it must arrange to call the main algorithm for each new sub-graph (this done in Step 6c).
CALL NEXT LEVEL—Step 6c: called only from “NO” result of Step 6a; makes multiple calls to Step 1a (see below).
For the current graph H, the k-level communities have been found in the form of B_{1}, B_{2}, . . . , B_{p′}. Since the algorithm has not reached the final h-level, the k+1-level communities must be constructed.
For each final k-level community B_{i }found by partitioning H by the current call of the main algorithm:
Output all B_{i}, each of which represents a level-h-community of G. This run of the main algorithm has terminated (though due to the recursive nature of the process, there may still be outstanding steps). The output may take any form, include printing, displaying on a display, recording to a memory or storage device or the output may be used as an input to another device or system or to control external devices or objects.
Example Processing of a Dataset
To see what partitioning by the number of connections with the above algorithm looks like, a simple example is useful. For this purpose, the dataset presented in
The dataset of
The steps below correspond to the same steps in FIGS. 7 and 8A-8C, and reference may be taken to these Figures, and the above explanations, to illuminate the following process. In the earlier Figures and text, step labels were abbreviated, such as S1 b, for Step 1b, and so forth.
Step 0: See
In step 0, the dataset of Graph G is input, resulting in the following:
Step 0a
Step 0b:
In Step 1, the main algorithm is started, a first maximal spanning forest is created, and the list of unsaturated edges is updated to indicate what edges remain after building this first forest.
Step 1a: find first maximal spanning forest.
Step 1b: update for the new forest
In step 3, a check is performed to see if all unused (not already in a potential new community) unsaturated (not used up by the spanning forests) edges are already assigned to a potential sub-cluster. If not, a new community loop is initialized.
Step 3a: check if all unsaturated edges are used
Step 3b: initialize new community loop
In step 4, the edge being processed is added to each tree (one in this case) to find what circuits are created.
Step 4-2b: Update the new Community
Step 4a: Determine if any Edges in the Potential New Community have not been Checked.
Step 4b: Community Complete
The process returns to step 3, a check is again performed to see if all unsaturated edges are already assigned to a potential sub-cluster. If not, a new community loop is initialized.
Step 3a: Check if all Unsaturated Edges are Used
Step 3b: initialize new community loop
Step 4-2b (first pass)
Step 4a Determine if any Edges in the Potential New Community have not been Checked.
Step 4c
Step 4-2a
Step 4-2b (Second Pass)
Step 4a
Step 4b
Step 3a
Step 3b
Step 4-2b
Step 4a
Step 4b
Step 3a
Step 6a
Step 6c: create new graph for next level.
Step 1a
Step 1b
Step 3a
Step 3b
Step 4-2b
Step 4a
Step 4c
Step 4-2a
Step 4-2b
Step 4a
Step 4c
Step 4-2a
Step 4-2b
Step 4a
Step 4b
Step 3a
Step 6a
Step 6c
Step 1a
Step 1b
Step 3a
Step 3b
Step 4-2b
Step 4a
Currently ƒ({ac, ab, bc})=[1,0,0], therefore e_{1}=ab has not been checked, resulting in a “YES”
Step 4c
Step 4-1a See
Step 4-1b
Step 3a
Step 3b
Step 4-2b
Step 6a
Step 6b
Step 6c
Step 6c (k=1 Continues)
Step 1a
Step 1b
Step 6a
Step 6c
Step 6c
Level 0 | Level 1 | Level 2 | Level 3 | ||
{a, b, c, d, e} | {a, b, c} | {a, b, c} | {a, b, c} | ||
{d, e} | {d} | {d} | |||
{e} | {e} | ||||
The above described algorithm shows the most basic and fundamental ideas of the processing approach according to the present invention. Following the same ideas, it can be further modified and improved for a lower complexity. Algorithm 2, shown in
Notation and Label System
In order to have a lower complexity, an alternative version has a more complicated data structure and the corresponding label system than the basic one.
The following is a list of a brief description of the data structure and the corresponding label system in this alternative algorithm.
1. The set B_{p }in the Algorithm of
Furthermore, for the purpose of reducing the complexity, the set B_{p }is created as a sequence (in Substep 4-1) so that the ordering of elements of B_{p }clearly indicates when and how those vertices were selected into the set B_{p}.
2. The adjustment sequences S(e) and I(e) will be created when they are needed (in the Substep 4-2) based on the information generated in Substep 4-1.
3. For the purpose of reducing complexity, the forests of T are all considered as rooted forests. An end-vertex of an unsaturated edge e_{0}is the root.
Main Program of Algorithm 2
Input: A graph G with w(e) as the multiplicity for each edge e and an integer h.
Goal: find all level-h communities in G.
Step 0. Start at k=0 (the program runs until k=h level is complete), H←G (a level-0 community), T←Ø (the set of spanning trees in H) and c_{T}(e)=0 for all e in E(G).
Step 1.
k←k+1.
(Note: if k>1, T={T_{1}, . . . ,T_{k−1}} is a set of edge-disjoint spanning trees of H, and c_{T}(e) is the coverage of each edge. These are outputs of Step 6 in the previous iteration of the main loop of the algorithm. When k=1, no spanning tree preexists).
Let E_{0}be the set of edges (unsaturated edges) e with c_{T}(e)<w(e).
Find a maximum spanning forest T_{0 }in H consisting of edges of E_{0}.
Let T←T+{T_{0}} and update the coverage c_{T}(e) and E_{0}as follows: c_{T}(e)←c_{T}(e)+1 if e is an edge of T_{0}, otherwise, c_{T}(e)=c_{T}(e) (no change), and delete all edges e in E_{0}such that c_{T}(e)=w(e).
Go to Step 2.
Step 2.
Let p←1 and go to Step 3.
Step 3. If E_{0}=Ø then go to Step 5.
Otherwise, go to Step 4.
Step 4.
Pick any e_{0}=xy ∈ E_{0}. Let B_{p}←b_{1}b_{2 }where bus x and b_{2}←y.
Let Q be the component of T_{0 }containing xy.
Let x be the root of each T_{i }(i>0) and Q.
Go to Substep 4-1.
Substep 4-1 (The Substep for searching of replaceable edges.) (See Subprogram 4-1 for detail.)
Outline and description. For each i∈ {0, 1, . . . , k−1}, and for each b_{j}∈ B_{p}, add all vertices of the directed path in T_{i }from b_{1}=x to b_{j }into B_{p}.
For every new vertex b_{j }added into B_{p}, always check: whether or not b_{j}∈ Q. If NOT, then stop the iteration of Substep 4-1 and go to Substep 4-2. Repeat this Substep until no new vertex can be added into B_{p}, then
H←H/B _{p} , E _{0} −E _{0} |H, p←p+1
and go to Step 3 (starting a new search for another level-k community). (Note, H/B_{p }is the graph obtained from H by identifying all vertices of B_{p }as a single vertex; E_{0}|H is the set of edges of e_{0}contained in the newly contracted graph H.) The new vertex of H created by contracting B_{p }is denoted by z_{p}.
Note, during the iteration, each new vertex b_{i }added into B_{p }is labeled with m(b_{i}), ε(b_{i}) and λ(b_{i}), which are to be used in creating the adjustment sequences S and I in Substep 4-2 in case that the spanning forest T_{0 }can be expanded. (For definitions of labels m, ε and λ, see Subprogram 4-2 for detail.)
Substep 4-2 (The Substep for expanding T_{0}.) (See Subprogram 4-2 for detail.) Outline and description. Create the adjustment sequences S(e) and I(e) based on the labels m(b_{i}), ε(b_{i}) and λ(b_{i}) generated in Substep 4-1 (See Subprogram 4-1).
Follow the adjustment sequences S and I to adjust and expand the forests of T.
And update the coverage c_{T }for the edge e_{0}.
Let B_{p}←Ø and erase labels m(b_{i}), ε(b_{i}) and λ(b_{i}) for all vertices of H.
Go to Step 3.
Step 5.
Let {v_{1}, . . . , v_{r}} be the vertex set of the resulting graph H (which has gone through a series of contractions in Step 4-1-2). Each vertex v_{i }is a child of H in the hierarchical tree, some of which are single vertices, while others represent non-trivial level-k communities.
If v_{i }is not a contracted vertex, then it is a child of H in the hierarchical tree, and no further action is needed for this vertex.
For a contracted vertex v_{i}=z_{p}, replace v_{i }with the corresponding community B_{p }and go to Step 6 for further iteration. Note that it is possible that some vertex of B_{p }is also a contracted vertex z_{p′}. In this case, all vertices of B_{p′} should be added into B_{p}. This procedure should be repeated for all possible contracted vertices in B_{p}.
Step 6.
If k=h, output all B_{i}, each of which induces a level-h community of G.
If k<h, then repeat Step 1 for H←G[B_{i}] for every i, T_{k}←T_{0}, and T={T_{1}|H, T_{2}|H, . . . ,T_{k}|H} which is a set of edge-disjoint spanning trees in H (as inputs of Step 1 for the next iteration of the algorithm at level (k+1)).
Subprogram 4-1: The Search of Replaceable Edges (Substep 4-1)
The subprogram in this subsection is the detailed Substep 4-1 of Algorithm 2.
Some notation and labels used in this subprogram are to be introduced:
For the sake of convenience, non-negative integers μ are represented by an ordered integer pair (α, β) where μ=αk+β with 0≦β≦k−1 and α≧0. In order to distinguish the different presentation of integer numbers, let M_{k }be the set of all those ordered integer pairs (that is M_{k}={0,1,2, . . . }×{0,1, . . . , k−1}=Z^{+}×Z_{k}).
B_{p}=b_{1}b_{2 }. . . b_{D} _{ e }is a sequence consisting of vertices of H that are already selected for the p-th potential community. b_{D} _{ e }is the last vertex of the sequence at the current stage.
m(b_{i})(∈ M_{k}) is an integer label of b_{i }∈ B_{p}: If m(b_{i})=(α, β), then the second component β of m(b_{i}) indicates that the vertex b_{i }is contained in the circuit of T_{β}+e where e is an edge joining two vertices b_{h}, b_{j }of B_{p }for some pair of indices h, j<i.
ε(b_{i})(∈ Z^{+}) is an integer label of b_{i }in B_{p}: b_{ε(b} _{ i } _{) }is a vertex in B_{p }and is also a child of b_{i }in a rooted tree T_{β} such that b_{i }is added into B_{p }because one of its children, b_{ε(b} _{ i } _{)}, was already in B_{p}. The edge b_{i}b_{ε(b} _{ i } _{) }is to be used for possible expansion of T_{0}.
λ(b_{i})(∈ Z^{+}) is also an integer label of b_{i}∈ B_{p}:
λ(b_{i})=min {j: b_{j}∈ D(T_{β}; b_{i})}
where m(b_{i})=(α, β) and D(T_{β}; b_{i}) is the set of all descendants of b_{i }in the rooted tree T_{β}. Then the vertex b_{i }is in the circuit of T_{β}+b_{h}b_{j }where h=λ(b_{i}) and b_{j }is not a descendant of b_{i}. Furthermore, j=ε(b_{h}) and b_{h}b_{j }is an edge contained in a tree T_{β′} where m(b_{h})=(α′, β′).
Labels m, ε and λ are to be used in Substep 4-2 for creating the adjustment sequences S and I.
There are some other auxiliary notation in the subprogram for the purpose of generating labels m, ε and λ, and the purpose of reduction of complexity.
Current status m_{C }(∈ M_{k}) is an indicator that indicates the current working status. At the initial situation, m_{C}=(0,0). When m_{C}=(α, β), the second component β indicates that the tree T_{β} is currently in the iteration of Substeps 4-1-3, 4-1-4, 4-1-5 and 4-1-6.
A “working zone” of Subprogram 4-1 is a subsequence b_{D} _{ s }, b_{D} _{ s } _{+1}, . . . , b_{D} _{ e }of B_{p}: D_{s}=min{j:m(b_{j})>m_{C}−(1,0)}.
The Substep 4-1-5 is to be processed along the working zone instead of entire sequence B_{p}. The use of “working zone” will eliminate some unnecessary search along the sequence B_{p }and therefore, reduce the complexity of the algorithm.
c(b_{i}): a (temporary) carry-on label for generating λ.
Subprogram 4-1 (The expansion of B_{p})
Subsubstep 4-1-1.
D_{s}←2, D_{e}←2,
m(x)=m(y)←(0,0),
and m_{C}←(0,1).
Subsubstep 4-1-2. (Check whether the expansion of B_{p }is ended.)
If m(b_{D} _{ e })≦m_{C}−(1,0) then
H←H/B_{p}, E_{0}←E_{0}|H, p←p+1
and go to Step 3 of Algorithm 2 (starting a new search for another level-k community). (Note, H/B_{p }is the graph obtained from H by identifying all vertices of B_{p }as a single vertex; E_{0}|H is the set of edges of E_{0}contained in the newly contracted graph H.) The new vertex of H created by contracting B_{p }is denoted by z_{p}.
Otherwise, go to Substep 4-1-3 and continue.
Subsubstep 4-1-3.
i←D_{s},
and let m_{C}=(α_{C}, β_{C}).
(In the rooted tree T_{β} _{ C }, all ancestors of vertices in the working zone will be added into the sequence B_{p }in Subsubstep 4-1-6.)
Subsubstep 4-1-4. (Update D_{s }for the next iteration in the tree T_{βC} _{+1}.)
If m(b_{i})<m_{C}−(0, k−2), then D_{s}←i, otherwise, D_{s }remains the same.
Continue.
Subsubstep 4-1-5. (Update c(b_{i}) if it does not exist.)
If c(b_{i}) does not exist, then
c(b_{i})←i.
Otherwise, do nothing.
Continue.
Subsubstep 4-1-6. (Adding vertices into B_{p }and labeling new vertices with λ)
Find the parent v of b_{i }in the rooted tree T_{β} _{ C }.
Case 1. v∉B_{p}. (This vertex v is to be added into B_{p}.)
Subcase 1-1. If v∉Q then the spanning forest T_{0 }is now ready for expansion (and the expansion of B_{p }stops): go to Substep 4-2 of Algorithm 2.
Subcase 1-2. If v∈ Q then this new vertex v is to be added at the end of the sequence B_{p }and all labels are to be updated for this new vertex as follows:
D_{e}←D_{e}+1, b_{D} _{ e }←v, λ(b_{D} _{ e })←c(b_{i}), c(b_{D} _{ e })←c(b_{i}), ε(b_{D} _{ e })←i.
And
i←+1
and go to Substep 4-1-4 (repeating for the next b_{i }in the sequence).
Case 2. v∈ B_{p}, say v=b_{j}, and j>i.
c(b_{j})←min{c(b_{j}), c(b_{i})}
if c(b_{i}) exists; or
c(b_{j})←c(b_{i})
if c(b_{i}) does not exist.
And
i←i+1
and go to Subsubstep 4-1-4.
Case 3. v∈ B_{p}, say v=b_{j}, and j<i.
Check whether b_{i }has reached the end of the working zone as follows.
If i=D_{e}, then
m_{C}←m_{C}+(0,1),
and erase all of “carry-on” label c, and go to Subsubstep 4-1-2.
If i<D_{e}, then
i←i+1
and go to Subsubstep 4-1-4.
Remarks about Subprogram 4-1
Fact. The label m of vertices in the sequence B_{p }form an non-decreasing sequence. That is,
m(b_{1})≦m(b_{2})≦ . . . ≦m(b_{D} _{ e }).
Fact. Whenever the Subsubstep 4-1-3 starts, the induced subgraph G[B_{p}] is connected, and, furthermore, B_{p }induces a connected subtree of T_{β} _{ C } _{−1}.
Fact. During Substep 4-1-6, those vertices b_{i }with m(b_{i})≦m_{C}−(1,0) induces a connected subtree of T_{β} _{ C }.
Fact. During Substep 4-1-6 Case 1, new vertices added into B_{p }are along a path in T from a pre-existing vertex by of B_{p }to the root b_{1}=x, where,
m_{C}−(0,k−1)≦m(b_{j})≦m_{C}−(0,1).
Fact. A vertex b_{i}∈ B_{p }with i≧3 is added into B_{p }because it is in the circuit of T_{β}+b_{h}b_{j }where h=λ(b_{i}), b_{h}is a descendant of b_{i }in the rooted tree T_{β} with the smallest subscript h, and b_{j }is not a descendant of b_{i }in the tree T_{β}. Furthermore, j=∈ (b_{h}) and b_{h}b_{j }is an edge contained in a tree T_{β′} where m(b_{h})=(α′,β′).
Subprogram 4-2: Expansion of T (for Substep 4-2, Expansion for T_{0})
At this stage, the inputs are (outputs of Substep 4-1-6, Subcase 1-1): m_{C}=(α_{C}, β_{C}), a vertex v∉Q and is the parent of b_{i}∈ B_{p }in the rooted tree T_{β} _{ C }.
Subsubstep 4-2-1. Let
b_{D} _{ e } _{+1}←v, λ(b_{D} _{ e } _{+1})←c(b_{i}), ∈ (b_{D} _{ e } _{+1})←i.
Subsubstep 4-2-2.
Set an edge-sequence S and an index-sequence I as follows:
S=(b_{i} _{ 1 }b_{i} _{ 1 } _{*}), (b_{i} _{ 2 }b_{i} _{ 2 } _{*}), . . . , (b_{i} _{ t }b_{i} _{ t } _{*});
I=β_{i} _{ 1 }, β_{i} _{ 2 }, . . . , β_{i} _{ t−1 }
where
i _{l}*=ε(b _{i} _{ l }), i _{l+1}=λ(b _{i} _{ l }) and m(b _{i} _{ l })=(α_{i} _{ l }, β_{i} _{ l })
for each l=1, . . . , t−1, and
b _{i} _{ 1 } =b _{D} _{ e } _{+1} , b _{i} _{ 1 } _{*} =b _{i}, and (b _{i} _{ l } b _{i} _{ l } _{*})=(b _{2} b _{1}).
(That is, for each l≦t−1, each b_{i} _{ t } _{* }is a child of b_{i} _{ l }in the rooted tree
each b_{i} _{ l+1 }is a descendent of b_{i} _{ l }in the rooted tree
with the smallest subscript i_{l+1 }in B_{p}.)
Subsubstep 4-2-3.
For each μ=(t−1), (t−2), . . . , 3, 2, 1 (note, in the reversed order):
And
T _{0} ←T _{0}+(b _{i} _{ 1 } b _{i} _{ 1 } _{*}).
Subsubstep 4-2-4. Update the coverage:
c_{T}←c_{T}(e)+1 if e=e_{0 }(and delete e_{0}from E_{0}if c_{T}(e_{0})=w(e_{0})), and c_{T}←c_{T}(e) otherwise. Erase labels: B_{p}, D_{s}, D_{e}, m, m_{C}, λ, ε, c and back to Step 3.
Remarks about Subprogram 4-2
In Subsubstep 4-2-2 (the construction of adjustment sequences S and I),
i_{1}>i_{1}*≧i_{2}>i_{2}*≧ . . . ≧i_{t}=2>i_{t}*=1.
And
for each μ=1, . . . , t−1, and is an edge contained in
Similar to the basic Algorithm (with the argument of replaceability and fixed frame of references), we are also able to show that each
in Subsubstep 4-2-3 remains as a tree/forest.
Additional Aspects of Invention
Various aspects of the present invention have been described above as an algorithm or mathematical approach. Embodiments of the present invention further include a method incorporating the algorithm as a step or steps, and further including input and output of data or results. The present invention also includes a computer program incorporating the algorithm or method, as well as a computing device running such a program and/or a processor programmed to perform the algorithm or method and/or output of data or results in a tangible form, such as on machine readable medium and/or use of such results to manipulate further devices or operations.
Embodiments of the present invention include a method of analysis of data sets, such as genomic data, social science data, pharmaceutical data, chemical data and other data, using a computer or processor to analyze the data, resulting in identification of relationships, as described herein, such as relatedness of genomic, social science or chemical data and/or output of such results.
The various patents, patent applications, publications and other references mentioned herein are incorporated herein, in their entirety, by reference, as if reproduced herein. Those of skill in the art will appreciate that these incorporated references may provide additional approaches to accomplishing some steps of the present invention and provide teaching to assist in practicing the present invention.
Further variations on the herein discussed embodiments of the present invention will be clear to those of skill in the art. Such variations fall within the scope and teaching of the present invention. It is the following claims, including all equivalents, which define the scope of the present invention.
Cited Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|
US5040133 | Jan 12, 1990 | Aug 13, 1991 | Hughes Aircraft Company | Adaptive clusterer |
US5263120 | Apr 29, 1991 | Nov 16, 1993 | Bickel Michael A | Adaptive fast fuzzy clustering system |
US5555196 | Jun 18, 1992 | Sep 10, 1996 | Toa Medical Electronics Co., Ltd. | Method of counting particles using degree of membership in clustering data points into subgroups |
US5703959 * | Jan 3, 1995 | Dec 30, 1997 | Toa Medical Electronics Co., Ltd. | Method and device for analyzing particles |
US5710460 * | Apr 21, 1995 | Jan 20, 1998 | International Business Machines Corporation | Structure for reducing microelectronic short circuits using spin-on glass as part of the interlayer dielectric |
US5745749 | Jun 21, 1995 | Apr 28, 1998 | International Business Machines Corp. | Method and system of file version clustering of object blocks using a compiler and database and having a predetermined value |
US5822218 * | Aug 27, 1996 | Oct 13, 1998 | Clemson University | Systems, methods and computer program products for prediction of defect-related failures in integrated circuits |
US5832182 * | Apr 24, 1996 | Nov 3, 1998 | Wisconsin Alumni Research Foundation | Method and system for data clustering for very large databases |
US5864845 | Jun 28, 1996 | Jan 26, 1999 | Siemens Corporate Research, Inc. | Facilitating world wide web searches utilizing a multiple search engine query clustering fusion strategy |
US5940832 * | Feb 27, 1995 | Aug 17, 1999 | Fujitsu Limited | Dynamic database structuring method and apparatus, and database clustering method and apparatus |
US5971596 * | Jun 14, 1996 | Oct 26, 1999 | Kabushiki Kaisha Toshiba | Constraint condition evaluation method and constraint condition evaluation system |
US6003029 | Aug 22, 1997 | Dec 14, 1999 | International Business Machines Corporation | Automatic subspace clustering of high dimensional data for data mining applications |
US6038557 * | Jan 26, 1998 | Mar 14, 2000 | Xerox Corporation | Method and apparatus for almost-constant-time clustering of arbitrary corpus subsets |
US6049797 | Apr 7, 1998 | Apr 11, 2000 | Lucent Technologies, Inc. | Method, apparatus and programmed medium for clustering databases with categorical attributes |
US6092072 | Apr 7, 1998 | Jul 18, 2000 | Lucent Technologies, Inc. | Programmed medium for clustering large databases |
US6134541 | Oct 31, 1997 | Oct 17, 2000 | International Business Machines Corporation | Searching multidimensional indexes using associated clustering and dimension reduction information |
US6195659 | Jul 14, 1998 | Feb 27, 2001 | Trw Inc. | Method and apparatus for morphological clustering having multiple dilation and erosion of switchable grid data cells |
US6269376 | Oct 26, 1998 | Jul 31, 2001 | International Business Machines Corporation | Method and system for clustering data in parallel in a distributed-memory multiprocessor system |
US6353832 | May 11, 1999 | Mar 5, 2002 | Lucent Technologies Inc | Selectivity estimation in spatial databases |
US6381605 | May 29, 1999 | Apr 30, 2002 | Oracle Corporation | Heirarchical indexing of multi-attribute data by sorting, dividing and storing subsets |
US6397166 | Nov 6, 1998 | May 28, 2002 | International Business Machines Corporation | Method and system for model-based clustering and signal-bearing medium for storing program of same |
US6466946 | Jun 7, 2000 | Oct 15, 2002 | Hewlett-Packard Company | Computer implemented scalable, incremental and parallel clustering based on divide and conquer |
US6487546 | Aug 27, 1998 | Nov 26, 2002 | Oracle Corporation | Apparatus and method for aggregate indexes |
US6505205 | Jan 3, 2002 | Jan 7, 2003 | Oracle Corporation | Relational database system for storing nodes of a hierarchical index of multi-dimensional data in a first module and metadata regarding the index in a second module |
US6584456 | Jun 19, 2000 | Jun 24, 2003 | International Business Machines Corporation | Model selection in machine learning with applications to document clustering |
US6640227 | Sep 5, 2000 | Oct 28, 2003 | Leonid Andreev | Unsupervised automated hierarchical data clustering based on simulation of a similarity matrix evolution |
US6643629 | Nov 18, 1999 | Nov 4, 2003 | Lucent Technologies Inc. | Method for identifying outliers in large data sets |
US6684177 | May 10, 2001 | Jan 27, 2004 | Hewlett-Packard Development Company, L.P. | Computer implemented scalable, incremental and parallel clustering based on weighted divide and conquer |
US6728715 | Mar 30, 2000 | Apr 27, 2004 | International Business Machines Corporation | Method and system for matching consumers to events employing content-based multicast routing using approximate groups |
US6751621 | May 2, 2000 | Jun 15, 2004 | Manning & Napier Information Services, Llc. | Construction of trainable semantic vectors and clustering, classification, and searching using trainable semantic vectors |
US6829561 * | Dec 9, 2002 | Dec 7, 2004 | International Business Machines Corporation | Method for determining a quality for a data clustering and data processing system |
US20020091667 * | Jul 11, 2001 | Jul 11, 2002 | Amit Jaipuria | Method and apparatus for optimizing networking potential using a secured system for an online community |
US20030224344 * | Mar 13, 2001 | Dec 4, 2003 | Ron Shamir | Method and system for clustering data |
US20040059521 * | Nov 7, 2002 | Mar 25, 2004 | Han Kyung Sook | Method for visualizing large-scale protein interaction data |
US20040267686 * | Jun 24, 2003 | Dec 30, 2004 | Jennifer Chayes | News group clustering based on cross-post graph |
US20050111386 | Nov 20, 2003 | May 26, 2005 | Kamal Jain | Packing steiner trees |
Reference | ||
---|---|---|
1 | Nash-Williams, C. St. J. A., J. London Math. Soc., 36 (1961) p. 445-450. | |
2 | Tutte, W.T., J. London Math. Soc., 36 (1961) p. 221-230. | |
3 | Zhang, C.Q.: "Integer Flows and Cycle Covers of Graphs," (1997) Marcel Dekker Inc., New York, ISBN 0-8247-9790-6. |
Citing Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|
US7774329 * | Dec 22, 2006 | Aug 10, 2010 | Amazon Technologies, Inc. | Cross-region data access in partitioned framework |
US8271468 | Jul 1, 2010 | Sep 18, 2012 | Amazon Technologies, Inc. | Cross-region data access in partitioned framework |
US8706782 | Jun 12, 2011 | Apr 22, 2014 | International Business Machines Corporation | Self-contained placement of data objects in a data storage system |
US8971665 | Jul 31, 2012 | Mar 3, 2015 | Hewlett-Packard Development Company, L.P. | Hierarchical cluster determination based on subgraph density |
US9298804 | Jul 11, 2014 | Mar 29, 2016 | Good Start Genetics, Inc. | Variant database |
US20100293206 * | May 12, 2009 | Nov 18, 2010 | Tatu Ylonen Oy Ltd | Clustering related objects during garbage collection |
US20160012149 * | Jul 9, 2014 | Jan 14, 2016 | International Business Machines Corporation | Using vertex self-information scores for vertices in an entity graph to determine whether to perform entity resolution on the vertices in the entity graph |
U.S. Classification | 1/1, 345/420, 706/20, 707/999.1, 707/999.103, 707/999.001, 707/999.004 |
International Classification | G06F17/00 |
Cooperative Classification | Y10S707/99944, Y10S707/99931, Y10S707/99934, G06K9/6224 |
European Classification | G06K9/62B1P2 |
Date | Code | Event | Description |
---|---|---|---|
Aug 8, 2006 | AS | Assignment | Owner name: WEST VIRGINIA UNIVERSITY RESEARCH CORPORATION, WES Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:ZHANG, CUN-QUAN;OU, YONGBIN;REEL/FRAME:018083/0048 Effective date: 20060503 |
May 7, 2008 | AS | Assignment | Owner name: WEST VIRGINIA UNIVERSITY, WEST VIRGINIA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:WEST VIRGINIA UNIVERSITY RESEARCH CORPORATION;REEL/FRAME:021091/0419 Effective date: 20080507 |
Jun 30, 2009 | CC | Certificate of correction | |
Jul 2, 2012 | FPAY | Fee payment | Year of fee payment: 4 |
Dec 2, 2016 | REMI | Maintenance fee reminder mailed |